The equipment inventory and AI assistant for the Cornell Tech MakerLAB. Students find a machine, read how to use it, and ask an assistant that answers from the lab's own records and manuals. Staff keep the inventory, work maintenance tickets and add new equipment in the same app.
Live: https://makerlab-ai.vercel.app
Deployed across the lab's ~100 machines, and the subject of an accepted demo paper at
ISAM 2026. ISAM 2026 demo: the abstract, figures and submission record are in
docs/isam-2026-demo/.
Tool pages. Fuzzy search and facets (category, material, training, location) across the catalogue; each tool has specs, location, required training and PPE, its manuals and SOPs, and a live table of individual units with status and condition. QR labels on the machines open the right page.
An assistant on every page. It answers from the live catalogue through tool calls, not from a prompt baked in at build time. Opened from a tool page it becomes a specialist on that machine: it searches the machine's manuals and cites the page it read, sees photos of error screens, replies in the language it is asked in, troubleshoots first and then files a maintenance ticket when something is genuinely broken.
Research and intake. Staff add equipment by pasting a list or a product link, or by photographing a label. The app identifies each item, researches it in the background (official name, description, manuals, a clean product photo) and leaves a draft for a person to approve. Existing tools can be re-researched the same way.
Admin. /admin holds the inventory editor, the maintenance and corrections queues,
project moderation, intake and research, user roles and titles, and an optional one-way
mirror into Notion. Everything created by the assistant or a student starts unpublished.
MCP access. The catalogue is a Model Context Protocol server at /api/mcp, so anyone
can query the lab from Claude, ChatGPT or another client — read-only without an account,
or as themselves after signing in. See docs/mcp.md.
Twelve languages, white-label. The interface is translated through next-intl;
branding, colours, institution and assistant name come from environment variables.
Next.js 16 (App Router, React Server Components, cacheComponents), React 19,
TypeScript, Tailwind CSS 4 with shadcn/ui. Postgres (Neon) through Drizzle ORM, with
pgvector for manual search; Vercel Blob for files (one public store, one private).
Better Auth with Google sign-in, restricted to Cornell addresses. Models through the
Vercel AI Gateway via the AI SDK; background jobs on the Workflow SDK. MCP via
@modelcontextprotocol/sdk. Vitest, React Testing Library, MSW and Playwright. Hosted on
Vercel.
Notion is no longer the data layer: it is read once by the import script, and optionally receives a one-way copy of the inventory.
git clone https://github.com/philosophercode/makerlab-tools.git
cd makerlab-tools
npm install
npm run dev # http://localhost:3000It runs with no configuration. With DATABASE_URL unset the app uses an in-process
Postgres (PGlite) seeded with two demo tools, shows a demo data banner, and stores
uploads in .blob-data/. Add AI_GATEWAY_API_KEY to .env.local to turn the assistant
on. .env.example documents every variable.
Working against the real inventory locally uses a persistent PGlite database:
PGLITE_DATA_DIR=.pglite-data npm run dev -- -p 3001 # set AUTH_BASE_URL=http://localhost:3001.pglite-data/ is single-process: stop the dev server before running an import,
db:migrate, manuals:index or data:push against it. How the local database is
filled from Notion and later copied to the hosted site is in
docs/deploy.md.
npm run dev # dev server
npm run build # db:migrate, then next build
npm run lint # eslint
npm run typecheck # tsc --noEmit
npm test # vitest: unit + integration + component
npm run test:e2e # playwright (once: npx playwright install chromium)
npm run test:all # all of the above; must pass before a merge
npm run spec:coverage # every route, tool, script and env var is documented
npm run eval # agent eval harness — real, paid model calls; never in CIThe test suite needs no credentials and makes no network calls.
| Path | What |
|---|---|
src/app |
Pages, /admin, API routes |
src/components |
UI: ui/ (shadcn), system/ (app primitives), chat/, admin/ … |
src/lib |
Data layer, action layer, capabilities, AI jobs, research, manuals, auth |
src/workflows |
Background jobs (Workflow SDK) |
messages/ |
Interface strings, 12 locales |
scripts/ |
Import, migration and maintenance tools, run with node --experimental-strip-types |
test/, e2e/, evals/ |
Shared test harness, Playwright specs, agent eval harness |
public/, data/ |
Bundled tool images; dated data bundles |
docs/ |
Everything below |
| Document | For |
|---|---|
docs/deploy.md |
Setting it up. Local stages, then a Vercel deployment step by step. |
docs/handover.md |
Running it. Accounts, routine tasks, backups, what to do when it breaks. |
docs/operations.md |
Monitoring, backups and restore. |
docs/architecture-guide.md |
How it works. Start here if you are inheriting the code. |
docs/assistant.md |
What the assistant can do, by role. |
docs/mcp.md |
Connecting Claude, ChatGPT, Codex and other MCP clients. |
docs/constitution.md |
The rules every change must respect. |
docs/specs/ |
Design specs and what is built against each. |
docs/MakerLab_design/DESIGN.md |
The design system. |
AGENTS.md, docs/architecture/ |
Conventions, then the app in detail area by area, for people and AI assistants alike. |
TESTING.md |
Test suite runbook. |
.env.example |
Every environment variable, with what it does. |
evals/README.md |
The agent eval harness. |
This repository once held two apps side by side. The older AirTable-backed version (v4)
was removed when the current app moved to the repository root; it is preserved at the
git tag v4-final.
Dated specs in docs/specs/ written before the move refer to paths under v5/.
Every feature starts with a spec that merges before the implementation
(docs/constitution.md, Article 1). Use
docs/specs/TEMPLATE.md, or /spec in Claude Code.
npm run test:all must pass before any merge.
This repository has no license file, so no open-source license applies and default copyright holds. Ask the owners below before reusing any part of it.
The Cornell Tech MakerLAB:
- Niti Parikh — Director
- Luis Rodrigo Navarro — Assistant Director
- Isaac Steinberg — Tech Lead